Abstract
<p>AI recommendations increasingly direct search, often optimized for satisfaction rather than value. In three preregistered experiments (N = 1,457 U.S. online and in-person participants), people searched a grid of hidden rewards while a simulated AI recommended either a reliably rewarding but lower expected-value region or one with a rare jackpot. Participants followed advice well above chance in both cases---even after rating the advisor unhelpful. The reliable recommendation produced more adherence, higher helpfulness ratings, and greater willingness to keep using it. Jackpot knowledge dampened but did not eliminate the pattern. Yet when shown the full reward landscape, most participants said an AI should recommend the jackpot region to others. A fourth study turned to the advisors: five large language models differed in whether they recommended the reliable or jackpot region. The results suggest AI advice sets the boundaries of search---and that satisfaction, not value, determines whether people keep following.</p>